[Priority 1] Add explicit scratchpad support for resumable analyses

[marknutter]

Problem

Currently, intermediate analysis results are stored in /tmp/rlm-sessions/ which is ephemeral and deleted on reboot. If a long-running RLM analysis crashes mid-dispatch, all progress is lost. There's no way to inspect what the agent was thinking during intermediate steps.

Proposal (from AIGNE paper analysis)

Create persistent scratchpad at /rlm/scratchpad/{session_id}/ that:

• Stores intermediate chunk analysis results during dispatch
• Captures LLM reasoning steps before synthesis (for debugging/provenance)
• Auto-prunes after session finalization (keep for 7 days)
• Selective promotion — ask agent 'should this scratchpad content become permanent memory?'

Implementation

1. Create scratchpad directory structure
2. Modify dispatch logic to save intermediate findings
3. Add CLI commands: rlm scratchpad list/view/promote/prune
4. Add auto-pruning cron job or hook

Impact

• Enables resumable long-running analyses
• Better debugging (see agent reasoning)
• Captures useful intermediate results

Effort

1-2 days

Related

Context Engineering Pipeline from 'Everything is Context' paper (arxiv 2512.05470)